FEFreeExamDumps.in

CCA-F Practice Questions — Page 12

Question 111

Open question ↗

Scenario: Prompt Engineering and Structured Output -

You must clarify an ambiguous transformation requirement to Claude. The team debates whether to write a long prose description of the desired behavior or to provide a few concrete input/output examples.

Which is most effective for communicating the expectation?

  • A.Provide 2–3 concrete input/output examples demonstrating the exact transformation.
  • B.Write a longer, more detailed prose description and avoid examples.
  • C.Increase temperature so the model explores interpretations.
  • D.Provide a single example and assume the model generalizes perfectly.

Question 112

Open question ↗

Scenario: Context Management and Reliability

A long support conversation is progressively summarized as it grows. By the end, the agent refers to the order amount as "about ninety dollars" instead of the exact $89.99, and the order date became "sometime in January."

What practice prevents this loss of precision?

  • A.Maintain a structured "case facts" block (IDs, exact amounts, dates) included in every prompt, independent of how history is summarized.
  • B.Disable summarization entirely and send the full raw history forever.
  • C.Ask the model to round all amounts to the nearest ten for consistency.
  • D.Lower the temperature so numbers don't drift.

Question 113

Open question ↗

Scenario: Context Management and Reliability

The `lookup_order` tool returns 40+ fields, but the current task needs only order_id, status, total, items, and return_eligible. Over many calls, the extra fields crowd the context window.

What is an effective mitigation?

  • A.A `PostToolUse` hook that trims the result down to only the relevant fields before the model sees it.
  • B.Tell the model in the prompt to ignore the fields it doesn't need.
  • C.Increase the context window by switching models.
  • D.Call the tool less often and cache full 40-field results.

Question 114

Open question ↗

Scenario: Context Management and Reliability

When aggregating a long report, key findings placed in the middle of a large block of text are sometimes missed by the model, while the same details at the top or bottom are reliably used.

Which phenomenon is this, and how do you mitigate it?

  • A.The lost-in-the-middle effect; place key findings at the start (and action items at the end) of the aggregated input.
  • B.Token starvation; simply raise `max_tokens`.
  • C.Schema drift; enforce a stricter JSON Schema.
  • D.Tool overload; reduce the number of tools.

Question 115

Open question ↗

Scenario: Context Management and Reliability

During a multi-day codebase investigation, the agent's context degrades and it starts referring to "typical patterns" instead of the specific classes it found earlier. You want findings to survive across context boundaries and new sessions.

What technique helps most?

  • A.Write key findings to a scratchpad file the agent can consult later instead of re-running discovery.
  • B.Keep everything in the conversation and never compact.
  • C.Ask the model to memorize the findings.
  • D.Restart from scratch each day with no notes.

Question 116

Open question ↗

Scenario: Context Management and Reliability

A search subagent catches a timeout internally and returns an empty result set marked as success. The coordinator concludes "no matches found" and moves on, unaware that the search actually failed.

What is the underlying anti-pattern?

  • A.Silent suppression: a failure was disguised as a valid empty result, so the coordinator cannot distinguish "no matches" from "search failed."
  • B.Excessive escalation of routine errors to the coordinator.
  • C.Infinite retries inside the subagent.
  • D.Over-narrow task decomposition by the coordinator.

Question 117

Open question ↗

Scenario: Context Management and Reliability

A document-extraction pipeline reports 97% overall accuracy and the team wants to fully automate it. A spot check reveals one document type (handwritten forms) has roughly 40% errors, hidden inside the aggregate.

What practice would have surfaced this risk?

  • A.Analyze accuracy by document type and field, and use stratified random sampling rather than relying on the aggregate metric.
  • B.Trust the 97% aggregate and automate everything.
  • C.Increase `max_tokens` for all extractions.
  • D.Ask the model to self-rate its confidence and trust it.

Question 118

Open question ↗

Scenario: Context Management and Reliability

A research synthesis states "The AI music market is estimated at $3.2B" with no source or date. Reviewers cannot verify it, and conflicting figures from other sources can't be reconciled.

How should claims be represented to preserve provenance?

  • A.Attach a structured claim→source mapping: source name, URL, publication date, and confidence for each claim.
  • B.Summarize all claims into a single fluent paragraph for readability.
  • C.Keep only the highest figure since it's likely most recent.
  • D.Drop any claim that has more than one possible source.

Question 119

Open question ↗

Scenario: Context Management and Reliability

Two sources report different values for the same metric: Source A (2023) says 10% and Source B (2024) says 15%. A naive synthesis labels this a "contradiction."

What is the correct handling?

  • A.Preserve both values with their dates and attribution; the difference may reflect change over time, not a contradiction—let the coordinator reconcile.
  • B.Average the two values to 12.5% and report a single number.
  • C.Pick the higher value because newer is always better.
  • D.Discard both values since they disagree.

Question 120

Open question ↗

Scenario: Context Management and Reliability

A support agent's escalation logic is being tuned. Designers debate using customer sentiment analysis and a model self-rated confidence score (1–10) to decide when to escalate to a human.

Which is the most reliable basis for escalation?

  • A.Explicit rule-based triggers (e.g., the customer asks for a manager, policy is silent on the request, or the agent cannot make progress), reinforced with few-shot examples.
  • B.Sentiment analysis, since upset customers always need a human.
  • C.The model's self-rated confidence (1–10), with automatic escalation below a threshold.
  • D.A trained classifier requiring historical labeled data, regardless of availability.